Paper supplementary material
Introduction
The supplementary material for the lancet paper looking at how different service components effect key clinical outcomes and health care use. Add in some text to talk about what we’re doing in the paper and what this document is going to show
Library
These are the packages that were used in the analysis of the data. Please note, the version numbers might be different, but the methods are still the same.
Data Linkage
Data linkage was carried out by the DaSH team at the University of Aberdeen. Several datasets were linked in order to provide the data necessary to complete the data analysis in this project.
ICD-10 Codes
Here is the list of ICD-10 codes used to detect Serious Infections, Cancer, and Cardiovascular events. The codes can be filtered in the table by typing out the key outcome of interest in the space provided.
Analysis
To demonstrate the analysis carried out in this paper, the results will be use to produce a synthetic dataset. The same model fitting procedures will then be shown alongside additional explanation and extra figures. For the demonstration. I’m only going to look at the Serious Infection data so as to avoid making this document too long.
Univariate
This analysis was concerned with looking at the effect of each service component by including each component in separate models adjusting for various demographic features and their 2-way interactions.
Figures for Coefficients
For each Service component, a Poisson regression was run while accounting for several demographic features. The results of these models can be seen in the panel below.
Local AAV pathway
Cohorted clinic
Joint/Parallel clinic
Vasculitis MDT
Local specialty/organ specific MDT only
Own day case unit
Specialist vasculitis nurse
Nurse led clinic
Wait time for new patients (<1 week)
Nurse Advice Line
Synthetic Data
Will probably need to do this twice for each dataset.
Multivariate Horseshoe
NB: I haven’t included the local specialty/organ specific as not everyone provided an answer to that question. So, might want to think about whether or not to include this information?
This section details the process of fitting the model with all service components included with their 2 way interactions only with the demographic information. This approach was opted for in order to estimate which features still appeared to have an impact on outcomes after accounting for the presence of the other service components. The horseshoe prior allows for coefficients to be close to 0 if there is little evidence the effect has much of an impact while allowing larger effects to move away from 0.
Should show a figure to show the specialties and how the responded to each question to show that we can look at these variables in this way.
From the below plot, we can also see that we can see that there are some variables that should only be entered once as they’re very strongly related. For example, Vasculitis MDT and Specialist vasculitis nurse are almost entirely the same. The same goes for Nurse led clinic and Nurse Advice line
Another way to look at similarity is the “jaccard” score. Effectively, this treats presence and absence asymmetrically so two absences aren’t used in this matching score. Generally this is used when there are a large number of items and 0’s a more likely to appear which inflates the similarity in a way that doesn’t matter in certain circumstances. Not sure whether the problem we’re looking at here needs this solution, but it’s worth a look I guess.
Try different grouping rules for nurses part
Has all Nurse values
Only missing one
Has at least one
Synthetic Data
Data Processing
Key Clinical Outcomes
In the above example, the synthetic data was already in a format that worked for the planned analysis. However, the real data had to be processed prior to being entered into the model. For example, inpatient hospital admissions can involve being transferred from one specialty to another. In these instances, a new row is created to represent this visit though it is still part of one continuous stay. This can cause issues for this type of analysis as the number of “relevant events” (in this instance Serious Infections that required hospitalisation) can become inflated. To avoid over-counting, a continuous stay marker can be generated to demonstrate whether a row represents part of a previous visit or a brand new admission.
Using this marker, each stay in hospital can be collapsed into one row with all unique ICD-10 codes appearing on the same row in the data. Additional processing involved removing events that occurred within some time window of a previous event to avoid counting the same event multiple times. In this study, a time window of 28 days was selected.
Emergency Health Care
As above, the Emergency care data could include multiple rows that all refer to a single hospital admission. For this analysis, only the first event within a continuous stay was included in the count. Therefore, only the initial reason for the admission was examined. An admission was counted as being an emergency if the admission type fell into one of several categories:
- 4: Deliberate Self Injury or Poisoning*
- 5: Road Traffic Accident*
- 6: Home Accident (includes Accidental Poisoning in the home)*
- 7: Other Injury (includes Accidental Poisoning other than in the home)*
- 8: Other (excludes Accidental Poisoning)*
- 30: Emergency Admission - no additional detail
- 31: Patient Injury - Self Inflicted (Injury or Poisoning)
- 32: Patient Injury - Road Traffic Accident
- 33: Patient Injury - Home Incident (incl. assault of accidental poisoning)
- 34: Patient Injury - Incident at Work (incl. assault of accidental poisoning)
- 35: Patient Injury - Other Injury (incl. assaly or accidental poisoning other than in the home or at work)
- 36: Patient Non-Injury (e.g. Stroke, MI, ruptured appendix)
- 38: Other Emergency Admission (including emergency transfers)
- 39: Emergency Admission - type not known
*NB: Some of the codes included are older codes as the data for this study ranged from 1996 to 2020.